Presented at the 1998 Workshop on Performance Analysis and Its Impact on Design

Characterization of MMX-enhanced DSP and Multimedia Applications on a General Purpose Processor

Ravi N. Bhargava, Ramesh Radhakrishman, Brian L. Evans and Lizy K. John,

Department of Electrical and Computer Engineering, Engineering Science Building, The University of Texas at Austin, Austin, TX 78712-1084 USA
ravib@ece.utexas.edu - radhakri@ece.utexas.edu - bevans@ece.utexas.edu - ljohn@ece.utexas.edu

Abstract

Proper use of native signal processing (NSP) instruction set enhancements can result in speedup for targeted applications. In this paper, we study the behavior of the X86 architecture's Multimedia Extension (MMX) instruction set on signal processing and multimedia algorithms and applications. In addition to quantifying speedup, we make comparisons based on detailed dynamic instruction profiling. We compare a suite of digital signal processing (DSP) and multimedia programs implemented in C code and the same programs implemented with calls to an MMX assembly library that performs filtering, vector arithmetic, and other relevant kernels. As expected, our analysis shows decreased execution time for most, but not all, of our MMX programs compared to their unmodified equivalents. The observed speedup for the programs using MMX ranges from 1.2 to 7.5. For each set of programs, we perform a detailed instruction level analysis using VTune. This allows us to isolate the specific reasons for speedup or lack thereof. This analysis allows one to understand which aspects of native signal processing are most useful and how it can be utilized most efficiently.

The full paper is available as PDF - Postscript - GNU-Compressed Postscript formats.

Last Updated 01/15/99.